Run Meta’s Muse Spark Open on a Mac Studio cluster, without buying one
Apple and Meta have shown that the 816-billion-parameter model runs on two to four M5 Ultra Mac Studios linked over Thunderbolt. We rack the Macs, wire the mesh, set up MLX and give you SSH access, all billed by the day.
Updated October 2026
On October 15, Meta released the weights for Muse Spark Open, an 816B mixture-of-experts model, and Apple published how to run it locally with MLX, splitting it across several Mac Studios connected by RDMA over Thunderbolt 5. According to CNBC, the required M5 Ultra Mac Studio with 256GB starts at $10,799, and you need at least two. maccloud rents you the same setup as a dedicated pod instead.
What it takes to run
Requirement
Muse Spark Open
Parameters
816B total, 42B active per token (mixture of experts)
Weights in BF16 (native)
More than 1.6TB
Weights in NVFP4 (4-bit)
About 493GB, too big for any single Mac
Hardware
2–4 Mac Studios with M5 Ultra
Interconnect
Thunderbolt 5 with RDMA (macOS 26.2+), MLX jaccl backend
Software
MLX + mlx-lm, launched across nodes with mlx.launch
A pod is two or four dedicated M5 Ultra Mac Studios racked side by side and connected in a full Thunderbolt 5 mesh. The rate is simply the daily rate of each Mac added together. Speeds below are Apple’s published numbers for these setups, not our own measurements.
2 × Mac Studio 512GB
Pod 2 · 1TB
2 × M5 Ultra · 1TB unified memory total · Pipeline or 2-way tensor parallel
$132/day
$118.80/day monthly rate (save 10%) · $3,564 per 30 days
Decode speed and time to first token at an 8,192-token context, from Apple’s MLX benchmark on M5 Ultra Mac Studios. Pods follow our Mac Studio launch schedule. We confirm a quote before anything is billed.
Rent or buy?
Buying four 256GB M5 Ultra Mac Studios at Apple’s starting price comes to at least $43,196, before Thunderbolt 5 cables, rack space, power and networking. A maccloud Pod 4 · 1TB costs $180/day at the monthly rate. That works out to about 239 days of rental before buying the same hardware would pay off, and you can stop at any time without having to resell four Macs.
Meta’s AI chief described the target user as developers who “tinker with it at small scale and then seek to ramp up.” Renting lets you test that first stage for a few days or weeks before you commit to buying hardware.
What we set up for you
Racked and meshed. Your Macs are installed next to each other, with each pair connected directly by two Thunderbolt 5 cables. This is the doubled-link setup Apple used, and it noticeably shortens time to first token on long prompts.
RDMA and MLX ready. We enable RDMA over Thunderbolt, then install MLX, mlx-lm and an mlx.launch hostfile listing your nodes, so a single command can start the model across the whole pod.
Yours to control. Every node gets its own public IPv4 and IPv6 address, an admin account with your SSH keys, rescue KVM access and A/B power. As on every maccloud machine, password SSH is disabled.
Weights on your account. You accept Meta’s license and download the model with your own Hugging Face token. The NVFP4 weights are about 493GB, which takes roughly seven minutes at the full speed of a 10Gbps port. Real downloads usually take longer.
Starting the model
These are the commands from Apple’s write-up, run from node 1:
# serve Muse Spark Open across all nodes (OpenAI-compatible, port 8080)
mlx.launch --hostfile hostfile.json --backend jaccl --env MLX_METAL_FAST_SYNCH=1 -- \
mlx_lm.server --model meta-models/Muse-Spark-Open-816B-A42B-NVFP4 --port 8080
# from your laptop: tunnel the endpoint over SSH, nothing exposed publicly
ssh -N -L 8080:127.0.0.1:8080 admin@node1.your-pod
Then point any OpenAI-compatible tool at http://127.0.0.1:8080/v1. That includes OpenCode, which Apple uses as its example, as well as coding agents, scripts and your own apps. Your prompts and code go only to your own Macs.
Coming soon: a managed endpoint
Our hosted API (one HTTPS endpoint and API keys in front of your own Macs, with no SSH tunnel needed) is in development Coming soon. Until it launches, the SSH tunnel above is the supported way to connect.
Questions
Why can't one 512GB Mac Studio run it?
Muse Spark Open has 816 billion parameters. Even quantized to 4-bit NVFP4, the weights take about 493GB, and inference needs working memory on top of that. Apple's own write-up says it does not fit on one machine, so it has to be split across at least two.
Can I use two 256GB Macs?
No. Two 256GB nodes give 512GB in total, which is not enough once macOS, the GPU memory limit and the context cache are accounted for. The smallest pod we offer is 2 × 512GB.
Is the pod shared with anyone?
No. Every Mac in a pod is dedicated to you, and the Thunderbolt links only connect your own nodes to each other. You get an admin account and SSH keys on every node.
Who handles the Meta license?
You do, the same way you would on your own hardware. Muse Spark Open is released under Meta's license, which allows limited commercial use. You accept it and download the weights with your own Hugging Face account. maccloud supplies the hardware and setup and doesn't resell access to the model.
Can I run other models on the pod?
Yes. A pod is just linked Macs you control. The same MLX setup runs other large open-weight models, and each node can run a smaller model on its own when you're not using the full cluster.
How is it billed?
Like every maccloud machine: a daily rate per Mac, drawn from prepaid credit, with 10% off when you activate for 30 days and 20% off for 365 days. A pod's rate is the sum of its Macs. We confirm the exact quote before you pay anything.
Request a Muse Spark Open pod
Choose a pod in the console and we’ll confirm availability and a quote within one business day. Nothing is billed until you approve it.
Muse Spark Open is a model by Meta, and Mac Studio, Thunderbolt and MLX are Apple trademarks or projects. maccloud is not affiliated with or endorsed by Meta or Apple. Performance figures are as published by Apple and will vary with context length, settings and software versions. Muse Spark Open is subject to Meta’s license terms.